ragleap-graph

Knowledge-graph-augmented retrieval for RAG systems — entity extraction, co-occurrence graphs, and graph-based document retrieval via Neo4j.

pip install ragleap-graph

Quickstart

from ragleap_graph import GraphConfig, GraphIndex

graph = GraphIndex(config=GraphConfig(
    uri="bolt://localhost:7687",
    user="neo4j",
    password="...",
))

graph.upsert_document(
    document_id="doc-1",
    title="Q3 Report",
    chunks=[{"text": "Acme Corp reported strong Q3 revenue growth."}],
)

docs = graph.find_documents_by_entities(["Acme Corp"])
related = graph.search_related_entities(["Acme Corp"], max_depth=2)

Current: v0.9.0 · 108 tests (92 passed without live credentials, 16 skipped)

Architecture

LLM-based extraction and dedup (v0.2.0+)

The default entity extraction is regex/heuristic-based (fast, free, zero dependencies). For messier input — e.g. inconsistent capitalization like "Acme Corp" vs "ACME Corp." — LLM-based extraction and dedup produce cleaner graphs. Requires the llm extra: pip install ragleap-graph[llm]

from ragleap.generation import ProviderConfig
from ragleap_graph import GraphConfig, GraphIndex, ExtractionConfig

graph = GraphIndex(
    config=GraphConfig(uri="bolt://localhost:7687", user="neo4j", password="..."),
    extraction=ExtractionConfig(
        method="llm",
        provider=ProviderConfig(provider="gemini", api_key="...", model="gemini-3.6-flash"),
        dedup_enabled=True,
    ),
)

Note: EntityDeduplicator merges spelling variants of an already-extracted name; it does not fix fragmentation caused by the regex extractor splitting one real-world entity into multiple candidates in the first place — see CHANGELOG.md for a real, measured example of this and how method="llm" avoids it at the source.

Cross-chunk relation extraction (v0.8.0+)

Relation extraction normally runs per-chunk, so a relation whose evidence spans two chunks — e.g. an entity named in chunk 1, referred to only by pronoun ("it", "the company") in a later chunk — can be missed. Enable cross_chunk_relations=True for one additional pass over the full document using every entity found across all chunks:

extraction=ExtractionConfig(
    method="llm",
    provider=ProviderConfig(provider="gemini", api_key="...", model="gemini-3.6-flash"),
    extract_relations=True,
    cross_chunk_relations=True,
)

Known limitation: this depends on the provider's reasoning ability to resolve the reference correctly. Live-verified working with Gemini; small local models (e.g. qwen2.5:0.5b) were found, via live testing, to produce an incorrect relation rather than none on this task — not recommended for this feature without independently verifying its output.

Ontology cross-validation (v0.9.0+)

Constrain which relation_type values are valid between which entity_type pairs. A relation violating the ontology is dropped (with a WARNING logged) rather than written to Neo4j; relation_type values not listed remain unconstrained. Requires entity_types= to also be set:

extraction=ExtractionConfig(
    method="llm",
    provider=ProviderConfig(provider="gemini", api_key="...", model="gemini-3.6-flash"),
    extract_relations=True,
    entity_types=["ORGANIZATION", "PERSON"],
    relation_ontology={"FOUNDED_BY": (["ORGANIZATION"], ["PERSON"])},
)

Operations

ragleap-graph is a knowledge-graph retrieval library, not a database administration tool. Backup and restore of the underlying Neo4j database are explicitly the operator's responsibility, not something this library wraps or automates -- see docs/operations/backup-and-restore.md for a concrete, tested procedure using Neo4j's native neo4j-admin tooling, and docs/adr/0001-backup-restore-ownership.md for the reasoning behind this scope decision.

Status

v0.9.0. Ported from a real production GraphService, adapted for standalone open-source use — see HANDOFF.md for the full design history. ragleap-rag >=0.12.0 is an optional dependency, required only for method="llm".

License

MIT